---
license: gemma
language:
- en
tags:
- agentic
- tool-calling
- function-calling
- code-agent
- gemma-4
- e4b
- lora
- unsloth
- sol-traces
- hermes-agent
base_model: unsloth/gemma-4-E4B-it
library_name: gguf
inference: false
---

# Gemma-4-E4B-Sol-Traces-v2

Continuation-trained coding-agent model from `unsloth/gemma-4-E4B-it`. Builds on the Sol-Traces v1 base with additional Hermes Agent session traces, expanding tool coverage from 5 to 99 tools and introducing real agent behavior patterns alongside the original deterministic trajectories.

Sol Traces denotes tool-use traces compiled from Hermes Agent session logs; the traces do not originate from OpenCode.

## Training Details

| Parameter | Value |
|---|---|
| Base model | `unsloth/gemma-4-E4B-it` (MoE, 4 active experts) |
| Base revision | `4e22d7e59e078e63a14f351efdc5232ed366b621` |
| Fine-tuning | LoRA continuation from v1 adapter (r=16, alpha=16, dropout=0) |
| Target modules | Language + attention only (k/q/v/o/gate/up/down projection) — 264 LoRA keys |
| Dataset | 21,438 train / 1,339 val / 2,534 test (merged v1-upgraded + v2-hermes-native) |
| Dataset provenance | `v1-upgraded-with-tool-responses + hermes-log-canonical` |
| Steps | 500 |
| Learning rate | 1e-4, cosine scheduler with 3% warmup |
| Batch size | 8 (1 × 8 gradient accumulation) |
| Max sequence | 8,192 tokens |
| Loss type | Assistant-only (tool responses excluded from loss) |
| GPU | Modal H100 80GB |
| Training time | ~2h 34min |
| Final train loss | **0.0255** |
| Validation loss | **0.0528** |
| Peak VRAM | 27.0 GiB / 80 GiB |

### Pilot (20-step verification run)

| Metric | Value |
|---|---|
| Training records | 264 (hermes-native canonical slice) |
| Steps | 22 |
| Training loss | 0.0102 |
| Eval loss | 2.133 |
| Runtime | 44.6s |
| Adapter integrity | 264 keys matched and loaded from v1 source ✅ |

## Dataset

The training dataset merges two sources:

### v1-upgraded (21,174 train / 1,324 val / 2,502 test)

The original Sol-Traces v1 corpus of 25,000 verified deterministic trajectories with full tool responses preserved and reformatted for the expanded Hermes-native tool schema. These are the same 224 repository-family trajectories from v1, re-rendered with complete tool-response pairs rather than the original tool-response-masked format.

### v2-hermes-native (264 train / 15 val / 32 test)

Redacted, verified Hermes Agent session traces drawn from `~/.hermes/state.db`. These trajectories use the full Hermes-native tool schema (99 tools) and reflect real agent behavior patterns including:

- **Diverse tool selection** — browser automation, file operations, MCP tools, modal/cloud APIs, delegate/spawn patterns
- **Evidence-grounded branching** — tool choices follow actual observation output, not predetermined reference paths
- **Multi-turn recovery** — retries after failed commands, alternative file discovery routes
- **No-change decisions** — correct identification that no code change is needed

### Combined tool registry

The full merged training uses a 99-tool schema drawn from the Hermes Agent runtime environment:

<details>
<summary>Full tool list (99 tools)</summary>

- `apply_learnings`, `apply_patch`, `autonomous_decide`, `background`
- `browser_click`, `browser_console`, `browser_fill_form`, `browser_get_images`
- `browser_press`, `browser_scroll`, `browser_snapshot`, `browser_type`, `browser_vision`
- `clarify`, `cost_check`, `cronjob`, `delegate_task`
- `evey_goals`, `execute_code`
- `fabric_brief`, `fabric_recall`, `fabric_search`, `fabric_write`
- `freeride free`
- `honcho_profile`, `honcho_search`
- `image_generate`
- `kill`, `learn_from_interaction`
- `list_files`
- `mcp__openrouter__generate_image`, `mcp__proxmox__*`, `mcp_chrome_devtools_*`
- `mcp_cloudflare_*`, `mcp_docker_*`, `mcp_insforge_*`, `mcp_leonardo_*`
- `mcp_porkbun_*`, `mcp_preference_*`
- `mem0_conclude`, `mem0_profile`, `mem0_search`, `memory`, `memory_decay`, `memory_score`
- `patch`, `process`
- `read_file`, `run_command`
- `search_files`, `send_message`, `session_search`
- `skill_manage`, `skill_view`, `skills_list`
- `task`, `terminal`, `todo`
- `tool_call`, `tool_describe`, `tool_search`
- `vision_analyze`, `watchdog_status`
- `web_search`, `write_file`
</details>

### Data provenance and privacy

| Guarantee | Status |
|---|---|
| Source logs | `~/.hermes/state.db` only |
| Secrets, credentials | Fully redacted: `[REDACTED]` |
| Private paths | Fully redacted |
| Session IDs | Opaque HMAC-derived identifiers only |
| Content consent | Authorized Hermes traces, last 60 days |
| Privacy post-scan | Zero findings |

## Files

| File | Size | Description |
|---|---|---|
| `gemma-4-e4b-sol-traces-v2-Q4_K_M.gguf` | ~5 GiB | Quantized merged model — recommended for deployment |
| `gemma-4-e4b-sol-traces-v2-f16.gguf` | ~14 GiB | Full F16 merged model — for custom quantization |
| `adapter/adapter_model.safetensors` | 35 MiB | LoRA adapter weights (for PEFT-based loading) |
| `adapter/adapter_config.json` | — | LoRA configuration (r=16, alpha=16) |
| `training_stats.json` | — | Full training metrics and run report |

> **Note:** The Q4_K_M file is the recommended deployment format for llama.cpp. The F16 is provided for downstream quantization experiments. The `adapter/` directory allows PEFT-based loading without merging.

## Usage (llama.cpp)

```bash
# Q4_K_M — one file, ready to go
llama-cli \
  -m gemma-4-e4b-sol-traces-v2-Q4_K_M.gguf \
  -ngl 99 \
  --prompt "Find all package.json files in the project"

# Server mode with tool support
llama-server \
  -m gemma-4-e4b-sol-traces-v2-Q4_K_M.gguf \
  -ngl 99 \
  -c 4096 \
  --host 127.0.0.1 --port 8096
```

## Usage (PEFT / Transformers)

```python
from unsloth import FastModel
from peft import PeftModel

base = "unsloth/gemma-4-E4B-it"
model, tokenizer = FastModel.from_pretrained(
    model_name=base, max_seq_length=8192,
    dtype=torch.bfloat16, load_in_4bit=False,
    token="hf_...",
)
model = PeftModel.from_pretrained(model, "./adapter/")
```

## What's new in v2

Sol-Traces v2 introduces two major improvements over v1:

### 1. Expanded tool schema (5 → 99 tools)

v1 restricted the model to 5 deterministic tools (`list_files`, `read_file`, `search_code`, `run_command`, `apply_patch`). v2 exposes the full Hermes Agent tool registry including browser automation (`browser_*`), MCP integrations (`mcp_*`), memory management (`mem0_*`, `memory`), task delegation (`delegate_task`), scheduling (`cronjob`), and cloud API access.

### 2. Real agent behavior traces

v1 trajectories were generated by a deterministic reference executor that always followed the same pattern: list → read → run → patch → verify. v2 includes real Hermes Agent session traces with genuine decision-making:

- **Branching tool selection**: The model sees examples of choosing between alternative tools for the same goal
- **Error recovery**: Trajectories where a command failed and the agent tried a different approach
- **No-change scenarios**: Examples where the correct response was to explain why no code change was needed
- **Multi-turn workflows**: Longer sequences involving browser interaction, API calls, and file operations

### Training approach

v2 uses continuation training from the v1 adapter rather than training from scratch:

1. Load the v1 r=16/alpha=16 LoRA adapter (264 keys verified)
2. Continue on the merged v1+v2 dataset for 500 steps (same LR, batch, scheduler)
3. Merge and export as F16/Q4_K_M GGUF

This preserves the reliable v1 behavior while adding the new v2 capabilities.

## Capabilities

The model excels at:

- **Function calling**: Selecting and populating the right tool from natural language (99-tool schema)
- **Code navigation**: Searching, reading, listing, and patching files in codebases
- **Shell execution**: Running commands with proper flags and paths
- **Browser automation**: Clicking, typing, scrolling, and taking screenshots of web pages
- **Task delegation**: Spawning sub-agents for parallel work
- **API integration**: Using MCP tools for cloud/docker/proxmox operations
- **Memory management**: Reading and writing persistent state through memory tools
- **Verification**: Running tests, checking outputs, validating results

## Comparison with Sol-Traces v1

| Metric | v1 | v2 | Δ |
|---|---|---|---|
| Training records | 21,174 | 21,438 | +264 |
| Tool schema | 5 (deterministic) | 99 (Hermes-native) | +94 |
| Training loss | 0.0096 | 0.0255 | +0.0159 |
| Eval loss | 0.0235 | 0.0528 | +0.0293 |
| Training time | 1h 03m | 2h 34m | +1h 31m |
| Data diversity | Narrow (2 tool sequences) | Broad (99 tools, real agent patterns) | Significant |

The higher loss numbers in v2 reflect the more diverse and challenging training distribution — the model is learning a much broader task space with less repetition, not regressing.

### v1 → v2 tool-routing baseline

| Tool | v1 Selection | v1 Exact Pass |
|---|---|---|
| `list_files` | 5/5 (100%) | 0/5 (0%) |
| `read_file` | 4/5 (80%) | 3/5 (60%) |
| `search_code` | 0/5 (0%) | 0/5 (0%) |
| `run_command` | 2/5 (40%) | 1/5 (20%) |
| `apply_patch` | 1/5 (20%) | 1/5 (20%) |
| `no-tool` | 4/5 (80%) | 4/5 (80%) |

The v1 E2B model showed a 30% overall routing pass rate (9/30). v2 routing evaluation results will be published when available.

## Training Stats

```json
{
  "status": "success",
  "run_kind": "e4b-v1-sol-traces-v2-full-continuation",
  "base_model": "unsloth/gemma-4-E4B-it",
  "base_revision": "4e22d7e59e078e63a14f351efdc5232ed366b621",
  "dataset_version": "sol-traces-v2.0.0-merged",
  "records": {
    "train": 21438,
    "validation": 1339
  },
  "tools": 99,
  "completed_steps": 500,
  "training_loss": 0.02548,
  "eval_loss": 0.05275,
  "learning_rate": 0.0001,
  "peak_memory_gib": 26.96,
  "runtime_seconds": 9260
}
```

## Comparison with Other Sol-Traces Models

| Model | Active Params | Q4 Size | Training Loss | Tools | Best For |
|---|---|---|---|---|---|
| **E2B v1** | ~5B | 3.2 GB | 0.0229 | 5 | Edge, CPU+GPU hybrid |
| **12B v1** | 12B | 6.8 GB | 0.0800 | 5 | Balanced performance |
| **E4B v1** | ~8B | 4.9 GB | 0.0096 | 5 | Best quality-size trade-off |
| **E4B v2** (this) | ~8B | ~5 GB | **0.0255** | **99** | **Full Hermes-native agent** |
| **26B-A4B v1** | ~8B* | 15.6 GB | 0.0113 | 5 | Maximum capability |

*E4B and 26B-A4B both activate 4 experts but have different base architectures (dedicated encoder vs unified).

## Limitations

- **Continuation-trained from v1**: The 500-step continuation is a targeted update, not a from-scratch training. Some v1 tool call patterns (e.g., `list_files` bias) may persist.
- **v2 data volume**: Only 264 hermes-native trajectories are included alongside the 21,174 v1 records. The v2 signal is small relative to the v1 base.
- **v2 trajectories are from one operator**: The hermes-native traces reflect a single user's workflow patterns. Broader diversity requires additional sources.
- **Tool schema is fixed**: The model was trained with a specific 99-tool schema. Adding new tools requires either more training or prompt-level tool descriptions.
- **Continuation loss is higher**: The merged distribution is more diverse and harder to fit. Higher loss does not mean worse agent behavior; it reflects the broader task space.
- **Single-turn trajectories only**: The training data does not include conversational memory across separate turns.
- **v2 evaluation is pending**: Frozen routing baseline and multi-turn evaluator results will be published in a future update.

## Disclaimer

**Use at your own risk.** This model is fine-tuned for coding-agent scenarios. The model owner accepts no liability for any damages or losses arising from its use. Users are responsible for compliance with applicable laws and regulations.