Gemma-4-E4B-Sol-Traces-v1

Repository coding-agent model fine-tuned from unsloth/gemma-4-E4B-it using LoRA on 25,000 verified deterministic reference trajectories.

The E4B run has the lowest reported final training loss among the four Sol-Traces v1 runs. Training loss is not an end-to-end tool-use evaluation; comparative benchmarks are not yet published.

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)
Fine-tuning	LoRA (r=16, alpha=16, dropout=0)
Target modules	Language + attention (k/q/v/o/gate/up/down projection)
Dataset	21,174 train / 1,324 val (gemma-4-native-tools format)
Dataset provenance	original-synthetic — 25,000 verified trajectories compiled from Hermes Agent session logs across 32,560 attempted scenarios
Epochs	1
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	~1h 03min (pilot 3min + full 1h)
Final train loss	0.00960
Validation loss	0.02347
Peak VRAM	27.0 GiB / 80 GiB
Throughput	6,518 tok/s

Dataset

The training dataset consists of 25,000 executable trajectories built by a deterministic scenario generator and replayed against generated repositories. It uses 224 language/task/variant repository families with repository-family-balanced splits:

• 21,174 training records
• 1,324 validation records
• 2,502 test records (see dataset_manifest.json)

Each trajectory is a full agent session containing:

• System instruction: Repository coding agent with tool-use guidelines
• User task: A well-scoped coding task from the deterministic fixture catalogue
• Assistant tool calls: Multi-step function-calling sequences using 5 tools:
  • list_files — glob-based file discovery
  • read_file — line-range file reading
  • search_code — regex code search (defined in the schema; not emitted by the v1 reference policy)
  • run_command — allowlisted shell execution
  • apply_patch — unified diff application
• Tool responses: Output, exit codes, truncation markers
• Verification: Post-task validation commands with pass/fail outcomes

Actual v1 task coverage

Type	Records
debugging	5,424
feature	4,709
refactoring	3,582
testing	3,607
build_config	3,269
integration	2,742
documentation_review	1,667

Repository fixtures cover TypeScript, JavaScript, Python, shell, configuration, Go, Rust, and JVM/Java.

Data generation and verification

Sol Traces are compiled from Hermes Agent session logs produced while running deterministic, seed-based coding scenarios through a reference executor. The scenarios define repository templates, task requirements, and verification commands; accepted records retain the corresponding tool-use events and verification outcomes. Records are included only when their configured post-task validation succeeds.

The v1 reference policy is intentionally narrow: it always lists files, reads the known implementation path, runs pre-patch verification, applies the reference patch, and reruns verification. search_code is included in the schema but has no v1 calls.

Key Statistics

Metric	Value
Trace source	Hermes Agent session logs (deterministic scenario generator + reference executor)
Attempted seeds	32,560
Accepted trajectories	25,000 (76.8% acceptance rate)
Rejections	5,872 structural duplicates + 316 verification failures
Provenance	original-synthetic
Repository families	224 language/task/variant families across 8 fixture categories

Files

File	Size	Description
gemma-4-e4b-sol-traces-v1-Q4_K_M.gguf	4.94 GiB	Quantized merged model (Q4_K_M) — recommended for deployment
gemma-4-e4b-sol-traces-v1-f16.gguf	13.92 GiB	Full F16 merged model — for custom quantization
training_stats.json	—	Full training metrics
dataset_manifest.json	—	Accepted-record counts, split ratios, and rejection summary

Note: The Q4_K_M file is the recommended deployment format. The F16 is provided for downstream quantization experiments.

Usage (llama.cpp)

# Q4_K_M — one file, ready to go
llama-cli \
  -m gemma-4-e4b-sol-traces-v1-Q4_K_M.gguf \
  -ngl 99 \
  --prompt "List the files in the repository matching *.py"

# With conversation template
llama-cli \
  -m gemma-4-e4b-sol-traces-v1-Q4_K_M.gguf \
  -ngl 99 \
  --temp 0.2 \
  --chat-template gemma \
  -p "Find all TODO comments in the codebase"

Capabilities

The model excels at:

• Function calling: Selecting and populating the right tool from natural language
• Code navigation: Searching, reading, and listing files to understand codebases
• Shell execution: Running commands with proper flags and paths
• Patch application: Making small, correct code changes via unified diffs
• Deterministic verification flow: Reproducing the fixture failure, applying the reference patch, and rerunning configured checks
• Verification: Running tests and validating changes

Comparison with Other Sol-Traces Models

Model	Active Params	Q4 Size	Training Loss	Speed	Best For
E2B	~5B	3.2 GB	0.0229	Fastest	Edge, CPU+GPU hybrid, low-resource
12B Unified	12B	6.8 GB	0.0800	Fast	Balanced performance
E4B (this)	~8B	4.9 GB	0.0096	Fast	Best quality-size trade-off
26B-A4B	~8B*	15.6 GB	0.0113	Moderate	Maximum capability

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

Why E4B?

The E4B achieved the lowest training loss (0.0096) of all four sol-traces models while maintaining a compact 4.9 GB Q4_K_M footprint. It offers:

• ~30% smaller than the 12B model (4.9 GB vs 6.8 GB Q4)
• ~3x smaller than the 26B model (4.9 GB vs 15.6 GB Q4)
• Highest throughput at 6,518 tok/s (1.4x faster than E2B)
• Best loss at 0.0096 (lowest across all four models)

This makes E4B the recommended default for most sol-traces deployments.

Limitations

• Fine-tuned for repository coding agent scenarios — general chat or creative writing may not benefit
• Single-turn trajectories only — no conversational memory across separate turns
• Tool schemas are fixed to the 5 tools in the training set
• Trained on synthetic trajectories — real-world coding patterns may differ

Training Stats

{
  "training_loss": 0.00960,
  "eval_loss": 0.02347,
  "steps": 377,
  "train_tokens": 24,704,714,
  "peak_vram_gib": 27.0,
  "throughput_tok_s": 6518,
  "runtime": "1h 03m"
}

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.